We assess all PMIP past1000 models for which SAT, SST and snow accumulation (precipitation - evaporation) are available, as well as the Community Earth System Model (CESM) Last Millennium Ensemble, for a total of twelve models, including eight PMIP3 models (MRI-CGCM3, MIROC-ESM, MPI-ESM-P, CSIRO-Mk3L-1-2, GISS-E2-R, BCC-CSM1-1, HadCM3 and the Community Climate System Model version4 (CCSM4)) , three PMIP4 models (MRI-ESM2-0, MIROC-ES2L and ACCESS-ESM1-5) , and the CESM-LME model . Additional PMIP past1000 models are excluded from this analysis due to our initial selection criteria (see Sect. 2.2). The resolutions and numbers of vertical layers for both the atmosphere and ocean are shown in Table .

The past1000 simulations serve to investigate the response to mainly natural forcing under background conditions resembling those of the current climate, i.e. the pre-industrial millennium. These simulations are based on a common protocol , describing a variety of suitable forcing boundary conditions, such as orbital parameters, solar irradiance, stratospheric aerosols of volcanic origins, and atmospheric greenhouse gas concentrations. The changes between the common protocol for PMIP3 and PMIP4 past1000 simulations are mostly derived from the use of newly available records, permitting a more comprehensive reconstruction of external forcing.

The CESM-LME employs version 1.1 of CESM with the Community Atmosphere Model version 5 (CESM1-CAM5) . The CESM-LME provides the largest ensemble of LM simulations with a single model to date, including a total of 13 members for the full forcing experiment. The only difference between ensemble members is a small (order 10-14) random roundoff difference in the air temperature field at the start of each simulation. The forcing over the LM includes orbital, solar, volcanic, changes in land use/land cover and greenhouse gas levels, and their implementation follows those used in PMIP3 .

Our knowledge of past Antarctic climate trends comes predominantly from a combination of proxy records from natural archives and paleoclimate models. To assess climate model performance, we rely on proxy records of Antarctica's climate and Southern Ocean conditions. We assess model skill by comparing the model outputs with four proxy-based reconstructions that are most relevant to the SMB: snow accumulation, surface air temperature (SAT), sea surface temperature (SST) and Niño 3.4 index. Other variables are also important for these processes but we are constrained by what reconstructions are available.

The past snow accumulation dataset is compiled by the PAGES Antarctica2K working group , which presents annual Antarctic snow accumulation variability at the regional scale over the past 1000 years. The dataset is comprised of 79 Antarctic ice core records, 44 of which cover the LM period (Fig. ). The estimates of snow accumulation are based on the physical distance between suitable age markers (bulk changes in isotopic composition reflecting glacial cycles, volcanic eruptions for decadal to millennial timescales, seasonal variations in stable water isotopes, and chemical species including sea salts, hydrogen peroxide, radio isotopes, and biologically controlled compounds within the ice core . These snow accumulation reconstructions provide valuable information on changes in certain regions; however, poor spatial coverage in some regions may result in misleading regionally averaged trends. suggest that a greater spatial representation with a higher number of ice core records, especially in the East Antarctic Plateau and Weddell Sea coastal regions, will improve the understanding of the true nature of Antarctic SMB in the past. We thus compare the reconstructed snow accumulation to GCMs output for each ice core record individually.

The surface air temperature (SAT) is obtained from a database compiled by the PAGES Antarctica2k working group . Paleotemperatures are reconstructed based on the statistical relationship between δ18O of water/precipitation and SAT. The database consists of 112 ice core records, shown in Fig. , which are temporally resolved at a 10-year average and reconstruct the last 2000 years. The reconstruction only provides regionally averaged LM SAT anomalies (referenced to the 1900–1990 CE period) time series over seven Antarctic regions: the East Antarctic Plateau (EAP), Wilkes Land coast (WL), Weddell Sea coast (WS), Antarctic Peninsula (AP), West AIS (WAIS), Victoria Land coast (VL), and Dronning Maud Land coast (DML) (Fig. ). In addition to these seven regions, there are also reconstructions for a continent-wide Antarctic region, broad-scale West and East Antarctic. Uncertainties arise from the uneven representation of ice core spatial coverage (EAP and WS) and the relative weak covariance on average between δ18O and SAT .

In addition to Antarctic climate records, we use reconstructions of Southern Ocean surface conditions. The PAGES Ocean2k group provides 57 SST reconstructions across the global ocean . Cores of sediment accumulated on the seafloor create excellent past archives and are used to reconstruct past ocean changes . Here, we focus on the four reconstructions located in the Southern Ocean (Fig. ). Of these, two reconstructions are annual, and the other two are seasonally averaged over the austral spring (SON). There are two types of proxies with their respective calibration: alkenones with the PRA1988 calibration and TEX86 with the KIM2008 calibration . The main uncertainty of those reconstructions is the relatively low temporal resolution with decade-scale gaps.

Teleconnections arising from the El Niño–Southern Oscillation (ENSO) play a crucial role in shaping recent Antarctic climate trends and SMB . ENSO exercises an influence on Antarctic climate by weakening or strengthening the Amundsen Sea Low, depending on its phase, which directly influences the atmospheric moisture over West Antarctica and, subsequently, the amount of precipitation . As a result, we include the ENSO index in the scoring. The ENSO index reconstruction is based on tree-ring data from Mexico and Texas, USA . A tree ring is a natural archive of past climate and has been widely used notably for its high temporal resolution and accuracy of dating . The dataset provides a reconstruction of Niño 1+2 (0–10° S, 90–80° W), 3 (5° N–5° S, 150–90° W), 3.4 (5° N–5° S, 170–120° W), and 4 (5° N–5° S, 160° E–150° W) indices over the past 700 years, extending back to 1300 CE, with the best verified portion beginning in 1400 CE. In this study, we focus on the Niño 3.4 reconstruction as it is the most commonly used index to define El Niño and La Niña events.

Overall the models show poor skill in simulating the snow accumulation over the AIS. Figure  shows the normalised regional snow accumulation score, calculated by averaging the score of all ice cores within each region. This highlights the spatial variability in snow accumulation and helps identify potential regional biases. The most consistent model across all regions is CSIRO-Mk3L-1-2 and is the best-scoring model in the AP and WS while maintaining a score of 3 or below everywhere else. CESM-LME mean is the best-scoring model in the EAP, VL and WAIS but scores in the bottom half of models in the WL. CCSM4 shows the same strengths and weaknesses as CESM-LME mean but performs slightly worse for all of them. MRI-ESM2-0 shows strength in the East Antarctic coastal region, as it is the best-scoring model in both the WL and DML and maintains a score of 5 or below everywhere else. ACCESS-ESM1-5 performs well in the WAIS and the EAP but scores in the bottom half of models in the WL. MPI-ESM-P performs relatively well for most regions but shows snow accumulation regional biases in the AP. MIROC-ES2L shows regional biases in the EAP, WAIS and VL as it consistently overestimates accumulation.

Figure  shows the mean, trends and temporal variability values of reconstructed and modelled time series for each ice core location allocated over their respective seven Antarctic regions. Details of the ice core records are shown in Table S1. Overall, the models show poor skill in simulating the snow accumulation over the AIS. They tend to overestimate the snow accumulation mean, while not capturing the trends and magnitudes of temporal variability. The greatest accumulation rates occur in the AP, WL and WAIS, while in the interior, VL and WS show modest accumulation rates.

Ten ice core records are located in the WAIS region. CESM-LME, MPI-ESM-P and ACCESS-ESM1-5 are models with mean accumulation within the reconstructed uncertainty for most proxies, while MIROC-ES2L, MIROC-ESM and MRI-CGCM3 consistently exhibit greater accumulation. No models succeed in capturing the signs and magnitudes of trends for locations with trends larger than 1.0 mm yr⁻². For locations with modest trends, models generally agree. CESM-LME performs best in terms of capturing temporal variability, while other models underestimate it.

Three ice core records are located in the AP region. CSIRO-Mk3l-1-2, MIROC-ES2l and BCC-CSM1-1 are always within the mean reconstructed uncertainty, and CSIRO-Mk3l-1-2 is the only model that captures the correct signs and magnitudes of trends for all three sites. MPI-ESM-P underperforms in this region and displays large differences in the trends and temporal variability.

Two ice core records are located in the WL region. Models underestimate the accumulation and fail to capture the significant trend of the second proxy, a recurrent issue in this study. The model that manages the best is MRI-ESM2-0, for both the mean and trends.

Only one proxy record exists for the WS. MIROC-ESM, MPI-ESM-P and BCC-CSM1-1 are the only models within the mean reconstructed uncertainty. Models agree with the modest trend but fail to be within the trend uncertainty and only CSIRO-Mk3L-1-2 captures the temporal variability. GISS-E2-R and MRI-CGCM3 exhibit the largest discrepancies in accumulation compared to the reconstruction.

Four ice core records are located in the VL region. CESM-LME scores the best for all three criteria. MIROC-ES2L shows potential regional bias as it displays the largest differences for mean, trends and temporal variability.

The EAP has the largest number of records with 21 ice core records, but most are located near the coast in close proximity to DML, making EAP poorly represented spatially. Proxies exhibit large annual temporal variations, suggesting that the annual accumulation rates vary substantially. This is not surprising considering that accumulation in this region is so low that even a small absolute increase in accumulation means a large relative increase . CESM-LME, GISS-E2-R, ACCESS-ESM1-5 and HadCM3 are the four models that best capture the mean. Models in general struggle to capture the trends, but ACCESS-ESM1-5 and MPI-ESM-P perform the best. Similarly, no models consistently reproduce the large temporal variabilities, but CESM-LME is consistently the closest. MIROC-ES2L exhibits a strong regional bias in the EAP as it shows the largest differences for all three criteria.

Because the time series only covers the last 100 years of the LM, the reconstructed uncertainties from the three ice core records in the DML, especially for the trends uncertainties, are much larger. This means that, despite apparent differences in sign and magnitude, most models fall within the uncertainty range of the reconstruction.

In contrast with snow accumulation, the models show agreement with the reconstruction in simulating SAT over the AIS. MRI-CGCM3 is the best-scoring model over five regions but is ranked last in the DML (Fig. ). While ACCESS-ESM1-5, GISS-E2-R and MIROC-ES2L are the top three scoring models for the DML, they score relatively poorly for the rest of the continent. It is important to note, however, that because the ice core records do not cover the full LM in the DML, this lack of temporal representation makes it difficult to rigorously assess the performance of the models in this region. Other noteworthy models are the CESM-LME mean, which is among the best-scoring models over six regions, and MPI-ESM-P, which scores in the bottom half of models only in the AP and DML. The warm bias of MIROC-ESM is reflected in its regional normalised score, as it is the worst-performing model in six regions out of seven and the second-worst in the seventh region. Despite its overall warm bias, MRI-ESM2-0 shows a better score than the overall model average over five regions (AP, WS, EAP, DML and WL), due to scoring better for the temporal variability criteria compared to the other models.

Figure  shows time series of the regionally averaged SAT anomalies for both the reconstruction and model simulations. Ice core reconstructions suggest a slight broad-scale cooling trend over most of continental Antarctica, with modest statistically significant temperature decreases over four regions: the EAP, WAIS, VL, and WL. The WS and AP do not display any statistically significant trends, while the DML shows the greatest temperature change. In contrast to these records, MIROC-ESM and MRI-ESM2-0 show positive trends for all regions, with the former starting to show temperature increases at the 1000 CE mark, suggesting that MIROC-ESM has a warm bias. For MRI-ESM2-0, the warm bias is more modest in magnitude. All other models are consistent with the general broad-scale cooling trend and generally show similar trend magnitudes.

In terms of the models consistent with the sign of change indicated by the reconstructions, only ACCESS-ESM1-5 and MRI-CGCM3 show a positive trend in the DML. Models show a slight cold bias in the WAIS and WL as the SAT anomalies are slightly lower in the models compared to the reconstructions. In each region, the main discrepancy between the models and the regionally averaged temperature reconstructions is the temporal variability, with modelled SAT exhibiting lower variability. Only in the WAIS region did the reconstruction also exhibit a similar low magnitude of variability.

The best-scoring model for the continent-wide Antarctica, West Antarctica (incorporating the AP and WAIS) and East Antarctica (incorporating the EAP, WL, WS, VL and DML) is MRI-CGCM3 (Fig. ). The second best-scoring model is CESM-LME mean. MPI-ESM-P performs slightly worse than when factoring in all seven regions (Fig. ) but still remains better than the average. MIROC-ESM is the worst-scoring model in terms of continent-wide Antarctica, West Antarctica and East Antarctica, followed by MRI-ESM2-0 with a normalised score of 6, a 4-point score difference.

Figure  is the same as Fig.  but for time series averaged for the continent-wide Antarctica, West Antarctica, and East Antarctica and shows a cooling trend in both continent-wide Antarctica and East Antarctica. Overall, the models show reasonable agreement with the reconstructions over these broader spatial scales. MIROC-ESM and MRI-ESM2-0 continue to show warm biases. Similar to the regional analysis of the seven individual Antarctic regions, models continue to show slightly colder SAT anomalies compared to the reconstructions.

Models show skills in capturing SST trends and temporal variability but exhibit a cool SST bias in all four locations. In Fig. , at site 1 (lat = -44.33°, long = -72.97°), the best-scoring model is MIROC-ES2L, followed by MRI-CGCM3 and GISS-E2-R. At site 3 (lat = -44.15°, long = -75.16°), the best-scoring models are ACCESS-ESM1-5, CSIRO-Mk3L-1-2 and GISS-E2-R. At site 4 (lat = -41°, long = -74.45°), the best-scoring model is MRI-CGCM3, followed by ACCESS-ESM1-5. MIROC-ESM is the worst-scoring model for all of the sites. For the second site (lat = -64.87°, long = -64.20°), off the western coast of the AP with SST seasonally averaged over the austral spring, all models disagree with the reconstruction and simulate SST at the freezing temperature. The models consistently simulate sea ice over that time of the year for this location, whereas the reconstruction implies the presence of sea ice only towards the end of the LM. Considering this model–proxy disagreement, we gave a score of 10 for all models.

SST reconstructions show a modest temperature cooling at all four marine sediment record sites (Fig. ). Only MRI-CGCM3 and ACCESS-ESM1-5 consistently show similar temperature means at sites 1, 3 and 4. CESM-LME mean, MPI-ESM-P and BCC-CSM1-1 display similar signs of change and magnitude of cooling, while MRI-ESM2 and MIROC-ESM are the only models that display warming trends at all sites.

Overall, there is a strong disparity between models in their ability to represent the Niño 3.4 index. MIROC-ES2L is the closest to the reconstruction at representing ENSO (Table ). It simulates a similar number of La Niña events and slightly fewer El Niño events than the reconstruction. CESM-LME mean and HadCM3 likewise capture well the number of La Niña events and slightly underestimates the number of El Niño events, followed closely by MRI-ESM2-0, which underestimates the number of La Niña events but shows a similar number of El Niño events to the reconstruction. The other models are scored more poorly with respect to ENSO. MPI-ESM-P, MRI-CGCM3, CCSM4 and CSIRO-Mk3L-1-2 produce a similar number of La Niña events to the reconstruction but differ considerably in terms of El Niño. BCC-CSM1-1, GISS-E2-R, MIROC-ESM and ACCESS-ESM1-5 differ from the reconstruction for the numbers of both La Niña and El Niño events.

Figure  shows the total score for each model along with their respective normalised snow accumulation, SAT, SST and Niño 3.4 index scores. As described in the method section, the best score is 1. The overall skill of the paleo-simulations for the LM is uneven depending on the variable considered, as no model performs equally well for all four climate variables. The mean score across the 11 models is 4.3. stated that models that score above the 90th percentile make up the subset of the best-scoring models. Only one model comprises this top 90th percentile – the CESM-LME mean with a score of 2.2. However, this is the mean score for CESM-LME as we averaged all 13 ensemble member scores together. The best PMIP past1000 model is CSIRO-Mk3L-1-2 with a score of 2.9. The poorest-performing models include MIROC-ESM and BCC-CSM1-1, two PMIP3 models, with respective scores of 8.2 and 6.3. The mean model score is 3.83 for PMIP4 models and 4.87 for the PMIP3 models. No PMIP4 models are part of the best-scoring models, but none are part of the poorest performing models (i.e. with an total score in the bottom half of models). All three PMIP4 models perform better than or equal to the mean of all of the models, whereas the PMIP3 models cover a more diverse range in scoring.

After evaluating models based on variables important for SMB over the LM, we now consider projections of AIS SMB. Similarly to , we defined the spatially integrated AIS SMB as precipitation minus sublimation. Here we compare the modelled AIS SMB projections between two scenarios, SSP2-4.5 and SSP5-8.5 (Fig. ). Of all the models we evaluated, future scenarios were not available for three models, CSIRO-Mk3L-1-2, MPI-ESM-P and HadCM3, and hence, we cannot examine their projected AIS SMB. The spatially integrated AIS SMB is projected to increase for the following 75 years (2025–2100) in both scenarios by all models. The spatially integrated AIS SMB values from the best-scoring model CESM-LME (CESM1-CAM5 with the LM forcing protocol) are projected to be 3107 ± 92 Gt yr⁻¹ (the associated uncertainties are ±1 σ) for SSP2-4.5 and 3521 ± 145 Gt yr⁻¹ for SSP5-8.5 from 2070 to 2100. For the same time period, AIS SMB from models scoring worse than the models mean (BCC-CSM1-1, GISS-E2-R, MIROC-ESM, and MIROC-ES2L) is projected to be 2992 ± 120 Gt yr⁻¹ for SSP2-4.5, and 3216 ± 148 Gt yr⁻¹ for SSP5-8.5, which is slightly lower than the best-scoring model.

In terms of trends, all models project positive SMB trends in all scenarios. For CESM1-CAM5, SMB is projected to have a mean trend of 5.2 ± 0.4 Gt yr⁻² for SSP2-4.5, and 13 ± 0.5 Gt yr⁻² for SSP5-8.5, while, in the worst-scoring models, the AIS SMB mean trend is projected to be at 2.5 ± 0.6 Gt yr⁻² for SSP2-4.5, and 6.8 ± 0.7 Gt yr⁻² for SSP5-8.5. The best-scoring model suggests stronger SMB means and trends with lower uncertainties for both scenarios compared with the worst-scoring models.

Figure  shows the projected regional changes in SMB over the AIS under the high-emission SSP5-8.5 scenario. Most of the future SMB changes are concentrated along the coast, with some smaller changes in the WAIS interior. The vast majority of coastal SMB changes are positive, with only a few models showing negative changes – notably GISS-E2-R in WL, MIROC-ES2L in VL, MIROC-ESM in WAIS and MRI-CGCM3 in WS. The ensemble mean only displays positive changes, with maxima in the AP, western WAIS, DML, and WL. The best-scoring model, CESM1-CAM5, projects positive changes along nearly all coastal regions except for VL and eastern WAIS, and shows strong agreement with ACCESS-ESM1-5 and CCSM4, two models that score slightly better than the model mean.

Here, we use the four best-scoring models overall (CESM-LME, CSIRO-Mk3L-1-2, MPI-ESM-P and MRI-ESM2-0) to investigate the relationship between climate variables during the LM. Figure a shows the precipitation anomalies relative to the historical period (1850–1900 CE). While there are some regional differences, particularly in coastal East Antarctica where CESM-LME is the only one that shows no notable anomalies, most of the best-scoring models show the greatest precipitation changes along the coasts. The largest discrepancy lies in the AP region, where both MRI-ESM2-0 and MPI-ESM-P simulate positive precipitation anomalies, while both CSIRO-Mk3L-1-2 and CESM-LME simulate negative precipitation anomalies.

For the relationship between temperature and precipitation (Fig. b), local precipitation sensitivities for all four best-scoring models are not uniformly distributed. All four models agree and show a consistent local-scale positive linear relationship and display similar regional sensitivity patterns. In general, models show lower sensitivities in the EAP and AP regions, ranging from 0 % °C⁻¹ to 12 % °C⁻¹ and higher sensitivities in the WS and VL regions with values that can reach up to 20 % °C⁻¹. CSIRO-Mk3L-1-2 is the only model that shows a stronger positive relationship in the interior.

Southern Ocean conditions exert a strong influence on Antarctic accumulation . To examine the relationship between ocean conditions and continental precipitation in the AP region, we look at the Pearson linear cross-correlation coefficients of modelled decadal SIC and precipitation (Fig. c) and of modelled decadal SST and precipitation (Fig. d). All four models agree and display similar correlation patterns in West Antarctica, with SIC and SST showing strong local spatial correlations with continental AP precipitation. The SIC-AP precipitation and the SST-AP precipitation correlations are slightly higher for MRI-ESM2-0 and CESM-LME than MPI-ESM-P and CSIRO-Mk3L-1-2. The models all exhibit high negative correlations between local SIC in the Bellingshausen and Weddell seas and AP precipitation. MRI-ESM2-0 and CESM-LME exhibit a high positive correlation between SIC in the Amundsen Sea and precipitation in the AP region, while MPI-ESM-P and CSIRO-Mk3L-1-2 exhibit a weaker one. For SSTs, here we show that models exhibit opposite correlations with high positive correlations between local SST in the Bellingshausen and Weddell seas and AP precipitation.

According to the four best-scoring models, regional precipitation patterns in West Antarctica are highly sensitive to local temperature and Southern Ocean condition (SIC and SST) changes. Even though the models disagree on the precipitation changes in the AP region, there is a consistent relationship between variables, as all four models do agree in the AP region in simulating a slightly positive linear relationship with local average warming, a strong negative correlation with local SIC and a strong positive correlation with local SST.

Both Antarctic ice core records and model simulations demonstrate clear regional differences during the LM. Some models are better at representing those regional features and some models show clear regional biases. However, several elements are remarkably consistent. To start, models tend to overestimate annual snow accumulation values everywhere except in the DML and WL regions. A recent evaluation of current and projected Antarctic precipitation in CMIP5 models has shown that compared with satellite data almost all the models overestimate current Antarctic precipitation, some by more than 100 % . This is a recurrent issue in current models that is likely due to poor representation of coastal topography which is a significant factor in how precipitation is represented for the AIS .

The snow accumulation trends provide a second example of mismatch between models and reconstructions. Models simulate the incorrect magnitudes of trends and for some, they also simulate the wrong sign. Furthermore, snow accumulation can be modulated by large-scale atmospheric circulation and ocean conditions. Regional and global modes of climate variability are suggested to be the dominant controls on regional climate in Antarctica during the LM . The mismatch in trends can be related to the model bias in simulating realistic patterns of decadal climate variability. Current generations of models struggle to simulate those features, especially in terms of their magnitude, spatial patterns and their sequential time development .

Model-simulated SAT agrees with the reconstructed SAT and shows generalised cooling over continental Antarctica but fails to reproduce the modest warming in the AP, and most models fail to reproduce the warming in the DML. Models that do reproduce the warming for both regions, MIROC-ESM and MIROC-ES2L, also exhibit a global warm bias. Therefore, their apparent agreement in these regions should not be interpreted as an indicator of good overall performance. found the overall skill in reconstructed surface temperature based on δ18O in the seven regions to be limited, but the reconstruction skill is higher and more uniform among reconstruction methods when the reconstruction targets are the bigger aggregated regions (West Antarctica, East Antarctica and Antarctica as a whole). Once averaged over these larger geographic areas, models show relatively strong agreement.

A final example of mismatch between models and reconstructions is the underestimation of Southern Ocean temperatures in models. Using SST reconstructions to constrain the model results is challenging, notably due to the large spatial gaps and temporal gaps in the records from the fact that few marine sedimentary archives have the resolution and age control necessary to reconstruct LM decadal-scale SST variability , and the potential for proxy-related biases . We only compare four site records with three of them located relatively close to one another. Hence, there are too few Southern Ocean records that cover LM to properly evaluate models on their ability to capture SST trends. This study would be more robust with an overall greater spatial coverage of proxy records.

Evaluations of PMIP and CESM LM simulations based on four different climate variables show that no model performs equally well for all variables. In general, models are better at simulating the SAT, and are substantially poorer at simulating snow accumulation as they have no skill in reproducing trends and temporal variabilities. LM has modest trends compared to other time periods ; our analysis suggests that capturing regional trends of such small magnitudes that we observe in the LM is still beyond current models' ability in view of their coarse resolution, among other limitations. Models show colder SST mean values but have skill in simulating trends and variabilities. Additionally, only a handful of models show skill in simulating ENSO . Nevertheless, some models are clearly better than others at capturing LM climate.

Atmospheric and oceanic horizontal resolution and the number of vertical layers vary widely among the models. We found that models with relatively high resolution for the atmosphere and ocean can perform equally as well as their coarse-resolution counterparts. There does not appear to be a clear relationship between horizontal resolution and model performance. There is also a lack of a clear relationship between vertical layers and model performance. For simulated snow accumulation, models participating in PMIP (but also CESM-LME) are run at insufficient resolution to provide accurate SMB estimates in the coastal regions . To resolve the SMB component characteristics in some of the narrowest coastal regions of Antarctica with complex topography, a grid spacing of about 50 km or finer is needed . There are two main modelling tools to improve on the relatively coarse resolution of GCMs: statistical and dynamical downscaling. Statistical downscaling is based on the use of statistical relationships between large-scale variables and local variables and has the advantage of being computationally efficient . Dynamical downscaling is based on the use of a RCM forced at its boundaries with GCM data . Both tools enable finer resolutions, potentially improving model performance.

There are many potential sources of model bias that are beyond the scope of this study. A non-exhaustive list includes the use of different cloud physics schemes, which can strongly influence precipitation over the ice sheet , the difficulty in simulating sea ice, which can affect moisture transport and precipitation on the continent , and the fact that models may not adequately represent weather patterns that could impact snow accumulation in certain regions of West Antarctica, potentially leading to mismatches with Holocene snow accumulation reconstructions . Such polar processes remain priorities for climate model development .

In this study, we compare two generations of paleo-simulations. The mean overall skill of PMIP4 models is greater than the mean overall skill of PMIP3 models, but there are more than twice as many PMIP3 models to analyse. For Antarctic climate during the historical period (1850–2000 CE), the latest generation of CMIP6 models has been shown to present no significant improvement at simulating some aspects of the modern climate with respect to CMIP5 models . Similarly, for past climates, within the few specific features we looked at, there seems to be little improvement between the different generations of models. It is possible that there are potential improvements in processes that we did not examine.

have evaluated CMIP5 and CMIP6 models over the historical period to look at which models capture the influence of anthropogenic warming on SMB. Although both studies share the same objective, our scoring method differs in that we consider multiple parameters important SMB as well as a longer timescale, and our model ensemble is also more limited because fewer models have run the past1000 simulations. Hence, it is useful to compare the results for models investigated by both studies. The two time periods have different forcing, which allows for a contrast between model responses to either anthropogenic or natural variability. GISS-E2-R and MPI-ESM-P are the best-scoring models in , while MIROC-ES2L and CCSM4 are the worst performing models. CESM1-CAM5 performed worse than their model average. For both studies, MPI-ESM-P, ACCESS-ESM1-5 and MRI-CGCM3 perform better than the model mean. Our results highlight that model evaluation studies should consider covering longer time periods for the full context of natural variability.

Previous studies have considered the historical time period to constrain future projections of Antarctic SMB. showed that models that best capture observed historical snowfall rates tended to project larger snowfall increase; although found a similar increase in SMB under the high-emission scenario, their subset of best-scoring historical models suggest smaller increases. Our sample size of available models is too small to be as conclusive as those studies, but with respect to the LM, the model that performs the best across multiple SMB-relevant variables projects greater future increases in AIS SMB. To strengthen the study, we encourage more climate modelling groups to participate in experiments such as past1000 so that projections can be constrained over timescales sufficient for end-of-century projections.

A positive SMB trend means mass gain over the surface of the ice sheet, a negative contribution to the global sea level . According to the best-scoring model, most of the SMB change at the year 2100 is concentrated along coastal regions, with maxima in the AP, western WAIS, DML and WL. The AIS interior remains relatively unchanged with the exception of small changes in the WAIS interior. However, at present, the ongoing dynamic ice loss in West Antarctica dominates the AIS mass balance . demonstrate that the increase in snowfall over the AIS during the 20th century did not offset the ocean-driven ice mass loss and only mitigates the AIS sea level rise contribution. Some studies have projected that ice discharge from West Antarctica will continue to dominate Antarctica's sea level contribution in the future even under low-emission scenarios . While some parts of the AIS will likely experience mass gain by the projected increase in SMB , this will likely not be enough to counteract the loss of mass from the marine basins of West and East Antarctica, even though the best-scoring model in our analysis shows an increasing SMB trend of 13 ± 0.5 Gt yr⁻² over the next century.

Although Antarctic SMB is projected to increase overall, there is still a question of how SMB changes may impact ice shelf stability in the future. discuss how atmospheric warming may lead to diverging SMB responses between grounded ice and lower-elevation ice shelves. Using an RCM to better represent changes in mass and energy fluxes at the surface, they find that the projected higher temperatures are likely to decrease SMB over ice shelves, mainly due to increased run-off and meltwater that can cause ice shelves to hydrofracture. Ice shelf collapse substantially increases the AIS sea level contribution in ice sheet model projections over the next 3 centuries . This highlights that regional downscaling of these coarser-resolution global models is essential for fully grasping the implications of these long-term SMB processes.

The strong AP precipitation increase, seen in two of the four best-scoring models (Fig. ), is in part attributed to the local atmospheric warming – MRI-ESM2-0 is the only model that simulates consistent warming in the AP during the LM period (Fig. ). Sea ice trends have an important influence on regional precipitation variations as sea ice-free and/or warmer SSTs promote evaporation, increasing the moisture content of the atmosphere and enhancing local precipitation . Hence, the potential warmer SSTs and/or SIC decline in the Bellingshausen and Weddell seas might have led to the increase in precipitation in the AP region in MRI-ESM2-0 and MPI-ESM-P models, while colder SSTs and/or greater SIC extent in the Bellingshausen and Weddell seas might have led to the decrease in precipitation in the AP region in the CSIRO-Mk3L-1-2 and CESM-LME models.

Sea ice trends can be driven by factors other than large-scale atmospheric circulation modes, but suggested that natural variability has played a crucial role over the last 2000 years, with the Southern Annular Mode (SAM) and ENSO believed to be driving regional climate heterogeneity for sea ice and sea surface temperature (SST) in the Southern Ocean. Those two modes wield their influence on West Antarctica by directly influencing the Amundsen Sea Low (ASL). Figure c and d can be interpreted as the ASL affecting the sea ice and precipitation rate for the AP. The model with the best score for AP snow accumulation (MIROC-ES2L) also has the best score for ENSO. CESM-LME also ranks highly in terms of ENSO and the AP snow accumulation. However, CSIRO-Mk3L-1-2 scores poorly in terms of ENSO but performs well with respect to AP snow accumulation.

Processes other than ENSO also impact how models simulate SAM and the ASL, which can impact precipitation change in the AP. While we do not investigate the impact of short-term fluctuations in precipitation in this study, this may contribute to the discrepancy in the AP region. Extreme precipitation events, such as atmospheric rivers (ARs), defined as long and narrow band of water vapour transport in the atmosphere , have been shown to impact the AIS mass balance . While occurring relatively rarely over Antarctica – only a few days per year – ARs transport large amounts of moisture from the mid- to high-latitudes and can lead to extreme precipitation and surface melt events . Model differences in the simulation of heat and moisture transport from mid-latitudes to the Antarctic continent could potentially be behind the discrepancy in the AP region, which is particularly impacted by these processes, and needs to be further assessed.